models.env

# ============================================================================
#  models.env  —  ONE place to choose every model the pipeline uses.
# ============================================================================
# `LiteRealitySettings` loads this before `.env`; environment and `.env` values win.
#
# Precedence: anything already set in `.env` or your shell WINS (each line uses
# `${VAR:-default}`), so this file is the default source of truth, not a lock.
# ----------------------------------------------------------------------------

# ── Agent harness — WHICH coding agent drives the file-editing sessions ─────
# The harness is the agent loop (file tools, capability tools, hooks, event stream); the model
# knobs below are the brain inside it. They are independent.
#   claude  — logged-in `claude` CLI via claude_agent_sdk (default). Full capability: capability
#             tools in-process, graceful step-budget wind-down, cost reporting, tool allowlist.
#   codex   — `codex exec` CLI. Capability tools bridged over stdio MCP; NO pre-tool hook (the
#             step budget becomes a hard stop), NO tool allowlist (shell is always available),
#             no cost reporting. Needs `codex` on PATH and a signed-in Codex account.
export LR_AGENT_PROVIDER="${LR_AGENT_PROVIDER:-claude}"

# Per-role overrides — EMPTY means "inherit LR_AGENT_PROVIDER". Set one to mix harnesses, e.g.
# author on claude and the procedural object agent on codex. (Left empty rather than
# `${...:-$LR_AGENT_PROVIDER}` because this file is read by python-dotenv, which does not expand
# a `$VAR` used as a default — it would arrive as the literal string.)
export LR_AUTHOR_PROVIDER="${LR_AUTHOR_PROVIDER:-}"          # Room.py authoring
export LR_MATERIALS_PROVIDER="${LR_MATERIALS_PROVIDER:-}"    # materials pass
export LR_QUALITY_PROVIDER="${LR_QUALITY_PROVIDER:-}"        # QC pass
export LR_REFINE_PROVIDER="${LR_REFINE_PROVIDER:-}"          # object refinement
export LR_PROCEDURAL_PROVIDER="${LR_PROCEDURAL_PROVIDER:-}"  # articulated GLB

# Codex-only knobs (ignored when the harness is claude). The model names below are Codex's, not
# Claude's — the HARNESS_MODEL family does not apply to a codex session. Leave the model unset to
# take whatever `codex` is configured to use.
export LR_CODEX_MODEL="${LR_CODEX_MODEL:-}"
export LR_CODEX_EFFORT="${LR_CODEX_EFFORT:-high}"   # minimal | low | medium | high

# ── Claude — reasoning / VLM (logged-in `claude` CLI, not metered) ──────────
# Options (any your `claude` CLI / provider can serve). Runs through the logged-in
# `claude` CLI (subscription, not per-token metered) — the $/MTok list prices below
# are for relative reference only:
#   claude-fable-5    — most capable overall; highest cost ($10/$50) [This is way too expensive and does not outperform claude-opus-5]
#   claude-opus-5     — current Opus; strongest agentic/coding tier ($5/$25) (default)
#   claude-opus-4-8   — prior Opus                            ($5/$25)
#   claude-opus-4-7   — older Opus                            ($5/$25)
#   claude-sonnet-5   — balanced, cheaper                     ($3/$15)
#   claude-haiku-4-5  — cheapest / fastest                    ($1/$5)

# The editor that authors & edits Room.py. Also drives classify, the VLM reader,
# and (unless overridden below) the critic. This is the main knob.
export HARNESS_MODEL="${HARNESS_MODEL:-claude-opus-5}"

# VLM critic. Defaults to HARNESS_MODEL — set only to run the critic on its own model.
export HARNESS_CRITIC_MODEL="${HARNESS_CRITIC_MODEL:-$HARNESS_MODEL}"

# Chair type-judge (frame/armrest/upholstery/base attributes during grouping).
# Same options as HARNESS_MODEL above — a stronger model here helps grouping accuracy.
export LR_CHAIR_JUDGE_MODEL="${LR_CHAIR_JUDGE_MODEL:-claude-opus-5}"

# Articulated-GLB procedural agent (doors/windows/articulated objects).
# Same options as HARNESS_MODEL above.
export LR_PROCEDURAL_MODEL="${LR_PROCEDURAL_MODEL:-claude-opus-5}"

# Completeness gate (reference-vs-render "is anything missing"). Runs once per build attempt per
# object and sits on the critical path. Defaults to HARNESS_MODEL. Lowering it is tempting but
# two-sided: too lenient ships objects missing parts, too strict costs a full agent rebuild.
export LR_COMPLETENESS_MODEL="${LR_COMPLETENESS_MODEL:-$HARNESS_MODEL}"

# ── OpenAI — reference image generation only ────────────────────────────────
# Options: gpt-image-2 (default, latest) · gpt-image-1 (legacy). Latest gives sharper, more
# faithful object renders; gpt-image-1 kept only as a fallback.
export LR_OPENAI_IMAGE_MODEL="${LR_OPENAI_IMAGE_MODEL:-gpt-image-2}"

# Which backend generates reference images: openai (default) or gemini. Naming a gemini-* model
# via --image-model switches on its own, so this is for pinning the choice across a whole run.
# Gemini needs $GEMINI_API_KEY and is roughly half the cost per reference (~$0.04 vs ~$0.08).
export LR_IMAGE_PROVIDER="${LR_IMAGE_PROVIDER:-openai}"
export LR_GEMINI_IMAGE_MODEL="${LR_GEMINI_IMAGE_MODEL:-gemini-2.5-flash-image}"

# ── Detection / embedding — HuggingFace checkpoints (local, CPU/GPU) ─────────
export LR_DINO_MODEL="${LR_DINO_MODEL:-IDEA-Research/grounding-dino-tiny}"    # GroundingDINO (detect)
export LR_DINO_EMBED_MODEL="${LR_DINO_EMBED_MODEL:-facebook/dinov2-small}"    # DINOv2 (grouping embeddings)